A fgr-am method and system for remote sensing scene recognition

Abstract

The invention discloses a FGR-AM method for remote sensing scene recognition, comprising: performing effective information enhancement processing and invalid information suppression processing on the image features extracted by the third and fifth bottleneck convolution modules; From the image features extracted by the first bottleneck convolution module, the contour information contained in the remote sensing image and the more interesting features in vision are extracted at the same time, and the detailed features contained in the remote sensing image are extracted from the image features extracted by the fifth bottleneck convolution module; Integrate channel attention and spatial attention enhancement features; map multi-dimensional features to orthogonal k-dimensional features to identify and classify remote sensing images. The present invention takes into account the main features and detailed features of the image, extracts and fuses the interested information and detailed information, so that the recognition accuracy of the network is improved, and the network can accurately identify the scene in complex scenes and similar scenes. identification.

Description

technical field [0001] The present invention relates to the technical field of computer vision, in particular to an FGR-AM method and system for remote sensing scene recognition. Background technique [0002] Remote sensing scene classification refers to dividing the image into blocks and labeling each block with an appropriate category (such as residential area, farmland, river, forest, etc.) according to the composition of the block. This has great significance for image management, retrieval, analysis, detection and identification of typical objects. As the resolution increases, images become more diverse, allowing fine-grained classification and recognition. At the same time, the details of high-resolution remote sensing images are more abundant, the features in the images are more diverse, and the objects on the ground are usually interlaced. The similarity between images of the same type decreases, while the difference between images of the same type increases signif...

AI Technical Summary

Problems solved by technology

[0005] However, these two methods are not suitable for dealing with high-resolution remote sensing images where there are scenes with high similarity, or there are two types of scenes with high similarity and large difference at the same time.

In fact, most of the current neural network methods for extracting remote sensing scene feature maps ignore the details when focusing on the main features of the image, or over-extract the details. The recognition accuracy decreases in similar scenes, which in turn makes the aforementioned problems difficult to solve

Images